recurrentNeuralNetwork.py 文件源码

def addInitialRNNLayer(mainGraph,
                       inputOperation=None,
                       activation=TanhActivation,
                       nHidden=100):
    """Add a RNN layer to input data

    Parameters
    ----------
    mainGraph : ga.Graph
        computation graph to which append the dense layer
    inputOperation : ga.Operation
        operation feeding the data to the layer, must have a shape of
        (nExamples, seriesLength, nFeatures)
    activation : ga.SingleInputOperation [class]
        activatin operation for hidden units
    nHidden : int
        number of hidden units

    Returns
    -------
    list(ga.Operation)
        List of activation operations from the RNN layer
    """

    nExamples, seriesLength, nFeatures = inputOperation.shape

    h0 = generateZeroVariable(shape=(nExamples, nHidden),
                              transpose=False)
    W = generateRandomVariable(shape=(nFeatures, nHidden),
                               transpose=False, nInputs=nFeatures * seriesLength ** 3)
    U = generateRandomVariable(shape=(nHidden, nHidden),
                               transpose=False, nInputs=nHidden * seriesLength ** 3)
    B = generateRandomVariable(shape=(1, nHidden),
                               transpose=False, nInputs=nHidden * seriesLength ** 3)

    h0op = mainGraph.addOperation(h0)
    Wop = mainGraph.addOperation(W, doGradient=True)
    Uop = mainGraph.addOperation(U, doGradient=True)
    Bop = mainGraph.addOperation(B, doGradient=True)

    hactivations = [h0op]

    # ------ append activation gates
    for indexRNN in range(seriesLength):
        xSliceop = mainGraph.addOperation(SliceOperation(inputOperation,
                                                         np.index_exp[:, indexRNN, :]))
        newHActiv = createRNNgate(mainGraph,
                                  xSliceop,
                                  hactivations[-1],
                                  Wop,
                                  Uop,
                                  Bop,
                                  activation)
        hactivations.append(newHActiv)

    return hactivations